Legal AI adoption: map the process before you pick the tool

A year after many legal departments and law firms rolled out their first AI tools, the question has shifted. Whether to use AI is mostly decided. What’s still open is where it belongs in the work, who checks what it produces, and how good individual results become something a whole team can rely on. In October 2026, a lawyer, the legal departments of the Thüga group and Austrian in-house lawyers arrived at the same answer independently: describe the process first, then decide for each step whether fixed logic, AI or a person handles it, and only then choose a tool.

What interests us is what these reports mean in practice. We put the three reports in context and use projects at the arbitration board SRUV, the law firm KUCERA Rechtsanwälte and ASD. DIGITAL SOLUTIONS to show how legal automation with AI delivers results that are not only faster but also reliable. The article closes with a self-check that helps you find the first process worth automating.

From “whether” to “how”: why the order matters now

For a long time, the debate about AI in legal work came down to a basic question: are we allowed to, do we want to, is it worth it? In the organisations that have now spoken publicly, that’s settled. Lawyers use AI for research, summaries and first drafts, and some already use it in production for standard contracts. That changes what a sensible first step looks like, because the goal now is for AI to actually save work without blurring who is legally responsible.

Teams that start with a tool end up looking for tasks that fit the tool. Teams that start with the process already know which steps eat time, where errors happen and where a person has to decide, so they can ask whether a tool covers exactly those points. The catch is that many processes in law firms and legal departments have never been written down.

One example from our own customers shows what that looks like. At KUCERA Rechtsanwälte, a mid-sized German law firm specialising in real estate law, the assistants checked what the firm called its “sacred deadline book”, a paper ledger, every morning for years, and once a week the whole team gathered while upcoming deadlines were read out from a list. That worked as long as there were enough people. When assistant positions became harder to fill, the ritual turned into a risk. The process existed, but only as a habit, and a habit can’t be handed over or automated.

Three voices, and what they mean in practice

All three sources were published in German. Quotes from them in this article are our own translations, rendered for meaning rather than word for word.

Control needs a fixed place in the process

In her interview with Anwaltspraxis Magazin, Hildegard Scholz describes how lawyers use AI today. Quick research is the most common use. What surprised her is that many treat AI as “a second opinion from a colleague”, used “not to lighten the load, but as a safeguard”. For her, the line is crossed where results are used directly and unchecked. Her advice to firms: “Don’t start with the system. Start with the process.”

What this means: As long as a lawyer uses AI on the side in a chat window as a second opinion, control depends on her personal discipline. She checks when she has time, and in her own way. Once the check becomes a fixed step in the process, it applies to everyone who uses that process, and it is clear who decides afterwards. Scholz’s distinction can be read as a division of labour: relief where rules apply, a safeguard where judgment is needed.

What it looks like in practice: KUCERA built its bots along exactly these lines, as digital assistants that take over small, time-consuming steps and leave the final review to the lawyers. Commercial register queries run automatically, document analysis delivers an instant summary, notarial client intake guides clients through structured data entry, and standard contracts such as leases are generated by a bot. The firm saves more than 20 percent of the time these tasks took and has produced over 1,000 documents this way. Matthias Fridriszik, attorney and legal engineer at the firm, says the recurring tasks now run as assistants in the background, “so we can get back to doing what we actually became lawyers for.”

Context has to be written down to apply to everyone

In its report of 13 October, Thüga describes how its legal department and colleagues from partner companies tested AI workspaces for a year. The first companies now use them in production. “The year of testing paid off,” says Dr. Winfried Rasbach, head of Thüga’s legal division. Lars Fischer, head of legal at badenova, adds: “With standard contracts in particular, it takes an enormous load off us.” Oliver Giesbert of Mainova, who works in legal and compliance management, calls AI “a sparring partner” but sees a clear limit: AI doesn’t understand how the company works, what the board wants or how much risk the organisation will accept. In his words: “The context is missing.” Jörg Gerbatsch, head of legal and insurance at TEAG, describes an internal rule that AI work products must be “both verified and labelled”.

What this means: “The context is missing” is often read as a weakness of the model. For an organisation, it is mainly a task. Risk appetite, thresholds, escalation paths and internal drafting standards that live only in individual heads can’t be passed to a model or to a new colleague. A better model won’t fix that. The same goes for Gerbatsch’s rule: “verify and label” is a process requirement. It only works once it is settled at which point the check happens, who does it and where it is recorded.

What it looks like in practice: The Arbitration Board for the Travel and Transport Sector (SRUV) had exactly this problem, at a scale of more than 45,000 cases in 2025. Every case was read and assigned by hand, which took 5 to 7 minutes, and the classification varied from one member of staff to the next. Together with Lexemo, SRUV translated its expertise into structured decision and rule chains, in other words it wrote down the context Giesbert says is missing. The bot reads the traveller’s complaint and the company’s response, separates disputed from undisputed points, assigns the case category and records a reasoning chain that shows how and why it decided. People stay involved in the steps that require evaluation. Preparing a case now takes about a minute, every case is classified by the same rules, and every decision step is documented. The reasoning chain does what the TEAG rule asks for: it stays visible which step ran automatically and on what basis.

Pain point first, and legal sets the frame

The write-up of the legal tech breakfast by Future-Law and VUJ gives the Austrian view. All participating in-house lawyers already use general-purpose language models. They agreed on the order: first clarify which pain points exist in the legal department, then look for the right tool for that task, and ask where generative AI is needed at all. According to the write-up, two more points came up. Many vendors now charge by usage, meaning by tokens, instead of per user. And it remained open how other business units can help themselves while the legal department sets the frame.

What this means: These look like two topics, budget and governance, but they are the same question. Which steps need AI, and which steps may another department run on its own? Both can only be answered from a written process. With usage-based pricing, every step that runs through a language model when a fixed rule would do costs money every single time it runs. And a business unit can only help itself safely if legal has defined the criteria beforehand.

What it looks like in practice: ASD. DIGITAL SOLUTIONS, the digital arm of the German commercial law firm Arnecke Sibeth Dabelstein, built a False Self-Employment Assessment Bot with its client Arcadis. It checks whether a contractor is in fact an employee under German law (Scheinselbstständigkeit). Legal teams used to do this by hand, over several hours, and quality depended on who did it. Now the bot asks structured questions, applies the legal criteria and returns a systematic assessment in seconds, cutting the time required by about 50 percent. The lawyers define the criteria and decision logic once, and the application runs at the client. That is the setup the Vienna discussion left open. Thomas Fischer, managing director of ASD (not to be confused with Lars Fischer of badenova), sums up the approach: “For us, automation is a product strategy, not an isolated tool experiment.” ASD’s own advice for getting started matches the Vienna consensus: start with concrete, recurring processes that have clear rules and evaluation criteria.

A framework for each step: four levels of automation

Kanzlei-IT-Portal, a German website on law firm IT, distinguishes four levels in its guide “KI in der Kanzleisoftware” (last updated 4 October 2026): Vorbereiten (preparing), Vorschlagen (suggesting), Nach Freigabe ausführen (executing after approval) and Regelbasiert automatisch (rule-based and fully automatic). The level should be chosen per type of task. The portal refers to guidance from the German Federal Bar (BRAK) that the duty of care rises with the degree of automation. The four levels are the portal’s own model, not a BRAK requirement.

Applied to the three projects, the model shows that one process usually combines several levels (this mapping is our own reading). KUCERA’s commercial register query is a clearly defined task and runs rule-based and automatically. At SRUV, the bot prepares the case and suggests a category, and the arbitrators decide. ASD’s False Self-Employment Assessment Bot applies fixed criteria and returns an assessment, while final legal responsibility stays with a person. So the level isn’t chosen for a tool or an organisation. It’s chosen for each step.

What the debate and practice show together

For e! by Lexemo, this is the point: the platform is built for exactly this groundwork. Lawyers draw their process on the Canvas as a decision tree, without code, and decide for each node whether it runs on fixed logic, uses AI or waits for a person. Logic steps run on rules, AI is used only where it is needed, and every decision step is documented. All three projects show that the experts build it themselves. After initial support from Lexemo, SRUV’s team has developed more than 400 automations on its own. At KUCERA, a three-person legal tech team led by a full-time legal engineer keeps finding new candidates and builds them without outside developers. And ASD turns legal and decision logic into client applications without writing code.

Self-check: which process is ready to automate?

This self-check helps you find out which process is the best first candidate for automation and what is still missing before it can be automated. Work through the five questions with the colleagues in your legal department or law firm who handle the process today, ideally before the first product demo is in the calendar.

1. Which pain point are you solving, and how often does it occur?

Name one task, not a category. “Contract work” is too broad. “First review of incoming NDAs from the sales team” is specific enough. Note how often it happens, who handles it today and where it regularly gets stuck. Scholz’s advice helps here: look for the steps that take a disproportionate amount of time and the points where time pressure puts quality at risk. You’ll also need the volume later for an honest estimate of usage-based costs.

2. Is the process written down?

Describe the process as it actually runs today, so that a new colleague could follow it without asking. Who starts it? What information and documents are needed? Which decisions are made along the way, and what happens when something is missing? Sticky notes on a wall are enough to begin with. You’ll often find that two experienced people handle the same process differently. At SRUV, exactly this inconsistency was one reason to write the decision logic down. Settle such differences before you automate anything.

3. For each step: logic, AI or a person?

Go through the process step by step. Deadlines, thresholds, mandatory fields and clear if-then decisions can usually run on fixed logic without any AI. AI helps where unstructured language has to be understood or a first draft written. A person decides where judgment, exceptions and responsibility are involved. The four levels from Kanzlei-IT-Portal help you decide, step by step, whether the system should prepare, suggest, act after approval or run on its own rules. Most teams find they need AI at fewer points than they expected.

4. Where is the context written down?

This is the gap Giesbert named. Write down what an experienced colleague knows and a model doesn’t: how much risk the organisation accepts, at what amount or contract term a matter is escalated, which clauses are never accepted, what the right wording looks like and who decides which question. As long as this context lives only in individual heads, quality depends on who happens to be in the office. Stored as a rule in the process, it applies to everyone, even when the expert isn’t there.

5. Who sets the frame, who uses it, and what is recorded?

Decide who owns the process, who may change it and who uses it. In companies this is often the Vienna setup: legal sets the frame, the business unit uses it. Make approved processes easy to find; KUCERA, for example, keeps a central SharePoint overview of all published bots. And decide what gets recorded: AI output that is verified and labelled under the TEAG rule, and approvals, for which Kanzlei-IT-Portal recommends a log of the step, the time and the approving person.

With answers to these five questions, the conversation with vendors changes. Instead of asking what a tool can do, you check whether it can represent your process as you defined it. The process description stays your knowledge, whichever tool you end up choosing.

Where to start

Today’s language models are good enough for many legal tasks. The bottleneck is the groundwork only lawyers can do: naming the problem precisely, laying the process open, deciding each step and recording the context that makes one organisation different from another. One process that comes up every week is enough to start. An afternoon, a wall and the colleagues who handle that process today are all you need for the first two steps.

Our Legal Use Case Canvas is a template for exactly this. On a single page, its eight fields take you from the problem through users, data sources and the process step where automation starts, to risks, governance and the success criterion. It’s free and comes with four filled-in examples from legal departments and law firms in Germany, Austria and Switzerland, including KYC/AML checks and NDA generation.

Frequently Asked Questions

Why do lawyers recommend starting with the process before choosing an AI tool?

Because a tool can only support work that has been defined. Before choosing AI or automation software, the team describes the process: which problem it solves, how often it occurs, which steps it has, who decides and what the result must be. Lawyer Hildegard Scholz put it plainly in Anwaltspraxis Magazin on 9 October 2026: “Don’t start with the system. Start with the process.” (our translation) Austrian in-house lawyers at the VUJ legal tech breakfast recommended the same order. With a written process, a team can check whether a product fits its work instead of bending the work around the product.

Does every step in a legal process need AI?

No. Deadline calculations, thresholds, mandatory fields, responsibilities and clear if-then decisions can run on fixed logic, which is more reliable and easier to inspect than a language model. AI suits steps where unstructured language has to be understood or a draft written. Judgment and final responsibility stay with a person. Because many vendors now charge by usage, this decision also affects the budget.

What do legal departments mean by “the context is missing”, and what helps?

The phrase comes from Oliver Giesbert (Mainova) in Thüga’s report of 13 October 2026. It means a model doesn’t know how the company works: risk appetite, escalation thresholds, non-negotiable clauses, corporate language, responsibilities. A better model doesn’t fix this. Written-down context, stored as rules in the process, does. SRUV translated its expertise into decision and rule chains so that more than 45,000 cases a year are classified by the same standards.

Are the four levels of automation a BRAK requirement?

No. The levels Vorbereiten (preparing), Vorschlagen (suggesting), Nach Freigabe ausführen (executing after approval) and Regelbasiert automatisch (rule-based and automatic) come from Kanzlei-IT-Portal’s guide “KI in der Kanzleisoftware” (last updated 4 October 2026). The portal refers to BRAK guidance that the duty of care rises with the degree of automation. The four-level model itself is the portal’s framework, useful for deciding how much automation fits each step of a process.

How can business units use legal processes without legal losing control?

By separating roles. The legal department describes and owns the process, sets rules, thresholds and escalation paths, and decides who may use it. The business unit applies it day to day. Cases above a defined threshold go back to legal, and every use is recorded. One example is ASD. DIGITAL SOLUTIONS’ False Self-Employment Assessment Bot: the lawyers defined the criteria, and the bot guides users at the client through the check.

How should AI output be documented?

The October 2026 reports name two building blocks. At TEAG, AI work products must be both verified and labelled. Kanzlei-IT-Portal recommends logging approvals with the step, the time and the approving person. Together they leave a clear trail: which step used AI, who checked it and who approved it when. This is good practice, but it doesn’t replace a professional-conduct review in the individual case.